#include "gromacs/domdec/partition.h"
#include "gromacs/essentialdynamics/edsam.h"
#include "gromacs/ewald/pme.h"
+#include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
#include "gromacs/ewald/pme_pp.h"
#include "gromacs/ewald/pme_pp_comm_gpu.h"
#include "gromacs/gmxlib/network.h"
*
* \param[in] step The step number, used for checking and printing
* \param[in] enerd The energy data; the non-bonded group energies need to be added to
- * enerd.term[F_EPOT] before calling this routine \param[in] inputrec The input record
+ * \c enerd.term[F_EPOT] before calling this routine
+ * \param[in] inputrec The input record
*/
static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
{
gmx_wallcycle* wcycle)
{
pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
- pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle, lambdaQ);
+ bool useGpuDirectComm = false;
+ gmx::PmeCoordinateReceiverGpu* pmeCoordinateReceiverGpu = nullptr;
+ pme_gpu_launch_spread(
+ pmedata, xReadyOnDevice, wcycle, lambdaQ, useGpuDirectComm, pmeCoordinateReceiverGpu);
}
/*! \brief Launch the FFT and gather stages of PME GPU
GMX_ASSERT(simulationWork.useGpuNonbonded,
"Can only offload buffer ops if nonbonded computation is also offloaded");
}
- flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
+ flags.useGpuXBufferOps = simulationWork.useGpuBufferOps && !flags.doNeighborSearch;
// on virial steps the CPU reduction path is taken
flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
const bool rankHasGpuPmeTask = simulationWork.useGpuPme && !simulationWork.haveSeparatePmeRank;
}
}
-/*! \brief Setup for the local and non-local GPU force reductions:
+/*! \brief Setup for the local GPU force reduction:
* reinitialization plus the registration of forces and dependencies.
*
- * \param [in] runScheduleWork Schedule workload flag structure
- * \param [in] cr Communication record object
- * \param [in] fr Force record object
+ * \param [in] runScheduleWork Schedule workload flag structure
+ * \param [in] nbv Non-bonded Verlet object
+ * \param [in] stateGpu GPU state propagator object
+ * \param [in] gpuForceReduction GPU force reduction object
+ * \param [in] pmePpCommGpu PME-PP GPU communication object
+ * \param [in] pmedata PME data object
+ * \param [in] dd Domain decomposition object
*/
-static void setupGpuForceReductions(gmx::MdrunScheduleWorkload* runScheduleWork,
- const t_commrec* cr,
- t_forcerec* fr)
+static void setupLocalGpuForceReduction(const gmx::MdrunScheduleWorkload* runScheduleWork,
+ const nonbonded_verlet_t* nbv,
+ gmx::StatePropagatorDataGpu* stateGpu,
+ gmx::GpuForceReduction* gpuForceReduction,
+ gmx::PmePpCommGpu* pmePpCommGpu,
+ const gmx_pme_t* pmedata,
+ const gmx_domdec_t* dd)
{
-
- nonbonded_verlet_t* nbv = fr->nbv.get();
- gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
+ GMX_ASSERT(!runScheduleWork->simulationWork.useMts,
+ "GPU force reduction is not compatible with MTS");
// (re-)initialize local GPU force reduction
const bool accumulate = runScheduleWork->domainWork.haveCpuLocalForceWork
|| runScheduleWork->simulationWork.havePpDomainDecomposition;
const int atomStart = 0;
- fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(
- stateGpu->getForces(),
- nbv->getNumAtoms(AtomLocality::Local),
- nbv->getGridIndices(),
- atomStart,
- accumulate,
- stateGpu->fReducedOnDevice(AtomLocality::Local));
+ gpuForceReduction->reinit(stateGpu->getForces(),
+ nbv->getNumAtoms(AtomLocality::Local),
+ nbv->getGridIndices(),
+ atomStart,
+ accumulate,
+ stateGpu->fReducedOnDevice(AtomLocality::Local));
// register forces and add dependencies
- fr->gpuForceReduction[gmx::AtomLocality::Local]->registerNbnxmForce(Nbnxm::gpu_get_f(nbv->gpu_nbv));
+ gpuForceReduction->registerNbnxmForce(Nbnxm::gpu_get_f(nbv->gpu_nbv));
- if (runScheduleWork->simulationWork.useGpuPme
- && (!runScheduleWork->simulationWork.haveSeparatePmeRank
- || runScheduleWork->simulationWork.useGpuPmePpCommunication))
- {
- DeviceBuffer<gmx::RVec> forcePtr =
- runScheduleWork->simulationWork.haveSeparatePmeRank
- ? fr->pmePpCommGpu->getGpuForceStagingPtr() // buffer received from other GPU
- : pme_gpu_get_device_f(fr->pmedata); // PME force buffer on same GPU
- fr->gpuForceReduction[gmx::AtomLocality::Local]->registerRvecForce(forcePtr);
+ DeviceBuffer<gmx::RVec> pmeForcePtr;
+ GpuEventSynchronizer* pmeSynchronizer = nullptr;
+ bool havePmeContribution = false;
- if (runScheduleWork->simulationWork.haveSeparatePmeRank)
+ if (runScheduleWork->simulationWork.useGpuPme && !runScheduleWork->simulationWork.haveSeparatePmeRank)
+ {
+ pmeForcePtr = pme_gpu_get_device_f(pmedata);
+ pmeSynchronizer = pme_gpu_get_f_ready_synchronizer(pmedata);
+ havePmeContribution = true;
+ }
+ else if (runScheduleWork->simulationWork.useGpuPmePpCommunication)
+ {
+ pmeForcePtr = pmePpCommGpu->getGpuForceStagingPtr();
+ if (GMX_THREAD_MPI)
{
- // PME force buffer on remote GPU -
- // event synchronizer received from other GPU only in case of thread-mpi
- if (GMX_THREAD_MPI)
- {
- GpuEventSynchronizer* const pmeSynchronizer =
- fr->pmePpCommGpu->getForcesReadySynchronizer();
- GMX_ASSERT(pmeSynchronizer != nullptr,
- "PME force ready cuda event should not be NULL");
- fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
- }
+ pmeSynchronizer = pmePpCommGpu->getForcesReadySynchronizer();
}
- else
+ havePmeContribution = true;
+ }
+
+ if (havePmeContribution)
+ {
+ gpuForceReduction->registerRvecForce(pmeForcePtr);
+ if (!runScheduleWork->simulationWork.useGpuPmePpCommunication || GMX_THREAD_MPI)
{
- // PME force buffer on same GPU - add dependency on PME force computation
- GpuEventSynchronizer* const pmeSynchronizer = pme_gpu_get_f_ready_synchronizer(fr->pmedata);
GMX_ASSERT(pmeSynchronizer != nullptr, "PME force ready cuda event should not be NULL");
- fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
+ gpuForceReduction->addDependency(pmeSynchronizer);
}
}
|| (runScheduleWork->simulationWork.havePpDomainDecomposition
&& !runScheduleWork->simulationWork.useGpuHaloExchange))
{
- fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
- stateGpu->fReadyOnDevice(AtomLocality::Local));
+ gpuForceReduction->addDependency(stateGpu->fReadyOnDevice(AtomLocality::Local));
}
if (runScheduleWork->simulationWork.useGpuHaloExchange)
{
- fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
- cr->dd->gpuHaloExchange[0][0]->getForcesReadyOnDeviceEvent());
+ gpuForceReduction->addDependency(dd->gpuHaloExchange[0][0]->getForcesReadyOnDeviceEvent());
}
+}
- if (runScheduleWork->simulationWork.havePpDomainDecomposition)
- {
- // (re-)initialize non-local GPU force reduction
- const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
- || runScheduleWork->domainWork.haveFreeEnergyWork;
- const int atomStart = dd_numHomeAtoms(*cr->dd);
- fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(
- stateGpu->getForces(),
- nbv->getNumAtoms(AtomLocality::NonLocal),
- nbv->getGridIndices(),
- atomStart,
- accumulate,
- stateGpu->fReducedOnDevice(AtomLocality::NonLocal));
+/*! \brief Setup for the non-local GPU force reduction:
+ * reinitialization plus the registration of forces and dependencies.
+ *
+ * \param [in] runScheduleWork Schedule workload flag structure
+ * \param [in] nbv Non-bonded Verlet object
+ * \param [in] stateGpu GPU state propagator object
+ * \param [in] gpuForceReduction GPU force reduction object
+ * \param [in] dd Domain decomposition object
+ */
+static void setupNonLocalGpuForceReduction(const gmx::MdrunScheduleWorkload* runScheduleWork,
+ const nonbonded_verlet_t* nbv,
+ gmx::StatePropagatorDataGpu* stateGpu,
+ gmx::GpuForceReduction* gpuForceReduction,
+ const gmx_domdec_t* dd)
+{
+ // (re-)initialize non-local GPU force reduction
+ const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
+ || runScheduleWork->domainWork.haveFreeEnergyWork;
+ const int atomStart = dd_numHomeAtoms(*dd);
+ gpuForceReduction->reinit(stateGpu->getForces(),
+ nbv->getNumAtoms(AtomLocality::NonLocal),
+ nbv->getGridIndices(),
+ atomStart,
+ accumulate,
+ stateGpu->fReducedOnDevice(AtomLocality::NonLocal));
- // register forces and add dependencies
- fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->registerNbnxmForce(
- Nbnxm::gpu_get_f(nbv->gpu_nbv));
+ // register forces and add dependencies
+ gpuForceReduction->registerNbnxmForce(Nbnxm::gpu_get_f(nbv->gpu_nbv));
- if (runScheduleWork->domainWork.haveNonLocalForceContribInCpuBuffer)
- {
- fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->addDependency(
- stateGpu->fReadyOnDevice(AtomLocality::NonLocal));
- }
+ if (runScheduleWork->domainWork.haveNonLocalForceContribInCpuBuffer)
+ {
+ gpuForceReduction->addDependency(stateGpu->fReadyOnDevice(AtomLocality::NonLocal));
}
}
haveCopiedXFromGpu = true;
}
+ if (stepWork.doNeighborSearch && ((stepWork.haveGpuPmeOnThisRank || simulationWork.useGpuBufferOps)))
+ {
+ // TODO refactor this to do_md, after partitioning.
+ stateGpu->reinit(mdatoms->homenr,
+ getLocalAtomCount(cr->dd, *mdatoms, simulationWork.havePpDomainDecomposition));
+ if (stepWork.haveGpuPmeOnThisRank)
+ {
+ // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
+ pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
+ }
+ }
+
// Coordinates on the device are needed if PME or BufferOps are offloaded.
// The local coordinates can be copied right away.
// NOTE: Consider moving this copy to right after they are updated and constrained,
// if the later is not offloaded.
if (stepWork.haveGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
{
- if (stepWork.doNeighborSearch)
- {
- // TODO refactor this to do_md, after partitioning.
- stateGpu->reinit(mdatoms->homenr,
- getLocalAtomCount(cr->dd, *mdatoms, simulationWork.havePpDomainDecomposition));
- if (stepWork.haveGpuPmeOnThisRank)
- {
- // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
- pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
- }
- }
// We need to copy coordinates when:
// 1. Update is not offloaded
// 2. The buffers were reinitialized on search step
{
GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
+ if (stepWork.doNeighborSearch)
+ {
+ /* On NS steps, we skip X buffer ops. So, unless we use PME or direct GPU
+ * communications, we don't wait for the coordinates on the device,
+ * and we must consume the event here.
+ * Issue #3988. */
+ const bool eventWillBeConsumedByGpuPme = stepWork.haveGpuPmeOnThisRank;
+ const bool eventWillBeConsumedByGpuPmePPComm =
+ (simulationWork.haveSeparatePmeRank && stepWork.computeSlowForces)
+ && pmeSendCoordinatesFromGpu;
+ if (!eventWillBeConsumedByGpuPme && !eventWillBeConsumedByGpuPmePPComm)
+ {
+ stateGpu->consumeCoordinatesCopiedToDeviceEvent(AtomLocality::Local);
+ }
+ }
}
}
wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSSearchLocal);
wallcycle_stop(wcycle, WallCycleCounter::NS);
- if (stepWork.useGpuXBufferOps)
+ if (simulationWork.useGpuBufferOps)
{
nbv->atomdata_init_copy_x_to_nbat_x_gpu();
}
if (simulationWork.useGpuBufferOps)
{
- setupGpuForceReductions(runScheduleWork, cr, fr);
+ setupLocalGpuForceReduction(runScheduleWork,
+ fr->nbv.get(),
+ stateGpu,
+ fr->gpuForceReduction[gmx::AtomLocality::Local].get(),
+ fr->pmePpCommGpu.get(),
+ fr->pmedata,
+ cr->dd);
+ if (runScheduleWork->simulationWork.havePpDomainDecomposition)
+ {
+ setupNonLocalGpuForceReduction(runScheduleWork,
+ fr->nbv.get(),
+ stateGpu,
+ fr->gpuForceReduction[gmx::AtomLocality::NonLocal].get(),
+ cr->dd);
+ }
}
}
else if (!EI_TPI(inputrec.eI) && stepWork.computeNonbondedForces)
wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
- if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
+ if (!stepWork.useGpuXBufferOps)
{
Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
}
if (simulationWork.useGpuNonbonded)
{
- if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
+ if (!stepWork.useGpuXBufferOps)
{
wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
if (!stepWork.useGpuFHalo)
{
+ /* We don't explicitly wait for the forces to be reduced on device,
+ * but wait for them to finish copying to CPU instead.
+ * So, we manually consume the event, see Issue #3988. */
+ stateGpu->consumeForcesReducedOnDeviceEvent(AtomLocality::NonLocal);
// copy from GPU input for dd_move_f()
stateGpu->copyForcesFromGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
AtomLocality::NonLocal);
|| (simulationWork.useGpuUpdate && haveDDAtomOrdering(*cr) && simulationWork.useCpuPmePpCommunication)
|| vsite)
{
+ if (stepWork.computeNonbondedForces)
+ {
+ /* We have previously issued force reduction on the GPU, but we will
+ * not use this event, instead relying on the stream being in-order.
+ * Issue #3988. */
+ stateGpu->consumeForcesReducedOnDeviceEvent(AtomLocality::Local);
+ }
stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
}
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff